Power Distribution Network Type Classification using a Machine Learning Approach

Abstract

Power distribution network vulnerability has been a critical component in measuring community resilience under natural disasters. Given overhead power lines exposed to extreme weather events are susceptible to large-scale damage and failure, it is imperative to identify if the power distribution network types are overhead or underground as part of the power outage prediction. As such data are not publicly available, we propose the application of machine learning techniques for power distribution network type classification. The purpose of this article is to improve the accuracy and generalizability of the power network type classification model proposed originally by ~\cite{Zhai2021}. Given that most power distribution networks follow road networks, we labeled the distribution network type for over 60,000 selected road locations across major cities in the United States. We then combine the power distribution network type dataset with nearby building characteristics, road types, probabilistic hazard maps, and geographical location information to form a complete dataset for the training of the network type classifier. We predict the network type at the building level, then aggregate the predictions back to individual road segments. We demonstrate the performance using different machine learning models, feature combinations, and aggregation methods. As a result, the best performance model is able to predict the existence of an overhead system with a testing accuracy of over 75\% and F1 score over 0.74. We conclude that our machine learning model is an effective and efficient tool for power distribution network type classification, which can be further applied to evaluate distribution network damage under natural disasters.